Case Study: Leveraging GenAI to Build AI-based Surrogates and Regressors for Modeling Radio Frequency Heating in Fusion Energy Science

E. Wes Bethel,Vianna Cramer,Alexander del Rio,Lothar Narins,Chris Pestano,Satvik Verma,Erick Arias,Nicola Bertelli,Talita Perciano,Syun'ichi Shiraiwa,Álvaro Sánchez Villar,Greg Wallace,John C. Wright

2024-09-10

Abstract:This work presents a detailed case study on using Generative AI (GenAI) to develop AI surrogates for simulation models in fusion energy research. The scope includes the methodology, implementation, and results of using GenAI to assist in model development and optimization, comparing these results with previous manually developed models.

Artificial Intelligence

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: in fusion energy science, traditional simulation codes (such as GENRAY) cannot meet the requirements of real - time experimental prediction and control because they are not fast enough. Specifically:

1. **Limitations of traditional simulation codes**:
- For example, the GENRAY/CQL3D simulation requires tens of minutes of running time without radial diffusion. Although this is acceptable for offline modeling, it is too slow for integrated modeling and real - time experimental control applications.

2. **Creating surrogate models using generative AI (GenAI)**:
- The paper aims to explore the use of generative AI to create numerical surrogate models that can replace high - fidelity physical simulations. These surrogate models can approximate the results of full - numerical physical calculations within a known or predictable error range.
- Through this method, researchers hope to significantly reduce the computing time, from tens of minutes to a few milliseconds (once the model is trained), thereby achieving faster real - time observation and control.

3. **Optimizing the model development process**:
- The research also explores how to use GenAI to assist in the entire model building, optimization, and evaluation process. Specifically, it includes:
- Providing initial code templates.
- Offering suggestions to optimize model parameters.
- Conducting more extensive model optimization using methods such as k - fold cross - validation.
- Exploring methods to reduce the number of input features (such as principal component analysis and correlation analysis) to reduce the computational cost of model training and inference.

4. **Comparison with manually - developed models**:
- The paper also compares the performance of AI - assisted code with previously manually - generated code in terms of model accuracy and computational requirements to evaluate the effectiveness of GenAI.

In summary, the main goal of this research is to accelerate the model development process in fusion energy science research and improve model performance and computational efficiency by introducing generative AI technology.